Z.ai opens GLM-5.3-Flash: mystery model Ox Alpha revealed, MIT-licensed 320B MoE
In one sentence On August 26 Z.ai released GLM-5.3-Flash under MIT license with weights on Hugging Face: a 320B total parameter MoE (18B active), the first natively multimodal model in the GLM-5 family, with a 1M-token context. It was the mysterious Ox Alpha running incognito on OpenRouter.
For a week the AI community puzzled over Ox Alpha, an anonymous model that appeared on the OpenCode and OpenRouter platforms with remarkable performance and no name attached. On August 26 the mystery closed: it was Z.ai GLM-5.3-Flash, now officially released with downloadable weights under the MIT license, the most permissive around.
The model uses a mixture-of-experts architecture: out of 320 billion total parameters it activates only 18 billion per response, like a large team of specialists where only the right ones work on each job. The result: small-model running costs with big-model knowledge. And it is the first in the GLM-5 family born multimodal, able to reason over images as well as text.
The price is the message: Z.ai claims ten times the cost efficiency of its predecessor, with API rates among the lowest on the market and a 50% launch discount until September 9. One detail that is not minor: during its incognito week the model ran entirely on AI chips manufactured in China, a practical demonstration of independence from the Western supply chain.
Companies
Z.ai
Tools
GLM-5.3-Flash
Tags
Sources